In recent years, Google has been on a mission to integrate generative AI into every conceivable product and initiative. This includes robots that summarize search results, interact with applications, and analyze data gathered from your phone. Often, the outputs generated by these AI systems can be unexpectedly impressive, even if they lack genuine understanding. But can they truly conduct scientific research? Google Research is now focused on developing AI that acts as a "co-scientist. " Their latest multi-agent AI system, built on the Gemini 2. 0 framework, targets biomedical researchers and is designed to assist by suggesting new hypotheses and research areas. However, this so-called AI co-scientist essentially functions like an advanced chatbot. A real scientist can utilize Google's co-scientist by entering their research objectives, concepts, and citations from previous studies, which allows the AI to propose new research directions. The system consists of various interconnected models that process the input data and tap into online resources to improve its suggestions. Within this framework, the different agents challenge one another, creating a "self-improving loop" akin to other reasoning AI models such as Gemini Flash Thinking and OpenAI's o3. Despite being a generative AI system like Gemini, it doesn't possess any truly new knowledge or ideas. Instead, it can make reasonable extrapolations from existing data. Ultimately, the AI co-scientist generates research proposals and hypotheses, and the human researcher can engage with the system via a chatbot interface to discuss these ideas. You might view the AI co-scientist as a sophisticated brainstorming tool. Just as individuals can share party-planning ideas with a consumer-level AI, scientists can generate new research concepts with an AI specifically designed for scientific inquiry. Testing AI in Science Currently, widely-used AI systems have a notorious issue with accuracy.

Generative AI tends to produce responses regardless of whether it has the right training data or model weights, and verifying facts using additional AI models doesn’t guarantee accuracy. With its reasoning capabilities, the AI co-scientist performs internal evaluations to enhance its outputs, and Google claims that these self-evaluation scores are linked to improved scientific accuracy. However, while internal metrics are informative, what do actual scientists think?Google asked human biomedical researchers to assess the proposals made by the robot, and they reportedly rated the AI co-scientist more favorably than other less specialized AI systems. The experts also noted that the outputs from the AI co-scientist displayed greater potential for innovative impact compared to standard AI models. That said, not all of the AI's suggestions are necessarily sound. Nevertheless, Google has collaborated with several universities to trial some of the AI-generated research proposals in laboratory settings. For instance, the AI recommended repurposing certain medications for the treatment of acute myeloid leukemia, and initial lab tests indicated that this approach was feasible. Research conducted at Stanford University also found that the AI co-scientist's treatment suggestions for liver fibrosis warranted further investigation. Although this research is certainly intriguing, referring to the system as a "co-scientist" may be somewhat overstated. Despite claims from AI leaders that we are approaching the advent of autonomous, thinking machines, AI is far from being capable of conducting scientific research independently. However, this AI co-scientist could still be instrumental in helping humans interpret and contextualize large datasets and research literature, even if it lacks true understanding or the ability to provide deep insights.